Alibaba turns Qwen into an autonomous GUI-clicking agent

If you have ever tried letting an AI agent build real software, you know the pain: it writes ten lines of code, hits a missing package, and halts until you paste the terminal logs back in.

Alibaba's Qwen team introduced Qwen 3.7 Plus, a multimodal model that wants to fix that loop by bringing visual perception, screen clicking, and command-line execution into one continuous workflow.

Eleven hours of non-stop coding

Instead of another chat demo, the team had Qwen 3.7 Plus build a full English vocabulary learning app completely autonomously.

The hybrid agent ran for over eleven hours straight. Across more than 1,000 internal tool calls, it generated over 10,000 lines of code without human hand-holding.

What makes this impressive is the breadth: the model drafted the technical spec, generated the code, installed dependencies, designed test cases, and actually watched the app GUI to run end-to-end user tests.

In another test, it autonomously operated the macOS Stocks app, decoded the visual hierarchy, and reconstructed the entire interface in SwiftUI before hooking it up to a real-time financial API.

There is also a browser extension called Qwen for Chrome. Give it permission, and it will navigate a complex cloud dashboard to locate the cheapest virtual server, configure storage and firewalls, and complete the checkout.

Built for GUIs, not pure science

The benchmark numbers show an intentional trade-off. Qwen 3.7 Plus is optimized for operating software rather than acing academic quizzes.

On screen-navigation benchmarks like AndroidWorld and ScreenSpot Pro, it comfortably outscores GPT-5.4, Opus 4.6 Max, and Gemini 3.1 Pro. Long-horizon task planning and terminal operations show a similar edge.

However, on heavyweight scientific reasoning benchmarks like MedXpertQA-MM, it still trails behind Gemini 3.1 Pro and GPT-5.4. Text capabilities are roughly on par with today's max-tier models without blowing them away.

To make adoption as frictionless as possible, Alibaba built native Anthropic API protocol support. You can drop Qwen 3.7 Plus straight into Claude Code or OpenClaw without rewriting your tool calls.

It also adds a preserve_thinking parameter to keep reasoning traces alive across conversational turns, which the team strongly advises for multi-step agent pipelines.

The model is proprietary with weights kept private, available solely via Alibaba Cloud Model Studio. Pricing is set aggressively at $0.40 per million input tokens and $2.40 per million output tokens, roughly six times cheaper on inputs than Qwen 3.7 Max.

Why it matters

This is one of the most pragmatic frontier plays we have seen out of China this year.

Running autonomous agents over hundreds of steps has traditionally been financially ruinous. If you point a Western frontier model at an eleven-hour coding task, the API invoice will make you wince. Qwen 3.7 Plus gives developers a dirt-cheap engine purpose-built for GUI clicks and terminal commands, with drop-in Anthropic compatibility.

Alibaba is effectively turning its cloud model into a blue-collar automation workhorse. If you are building automated UI testing, browser workflows, or routine app migration, this model offers immediate utility today.

The catch is architectural intelligence. Because it still lags top-tier models on dense academic and complex domain reasoning, you probably should not trust it to architect mission-critical infrastructure from scratch without review.

Still, by ignoring broad benchmark bragging rights to dominate the mechanical labor of clicking buttons and writing scripts, Alibaba just gave the entire agent ecosystem a very compelling reason to switch.